Field Snapshot

• LoRaWAN nodes log volumetric water content (VWC) at 6, 18, 30 in depths plus weather & irrigation telemetry
• Replace the placeholder with your field photo when available – visual context matters for stakeholders

Coverage Check

• 80% coverage on core soil & weather channels

• Stress indices and categorical annotations (crop, stage) too sparse for modeling this season

Raw Signal Behaviour

• Hourly VWC oscillates with sensor noise and irrigation pulses
• Daily averaging + Savitzky–Golay smooth preserves irrigation signature without dulling peaks

Domain Clean-up Moves

• PCHIP interpolation retraces the natural recharge/decay curve
• 15% spike flags isolate true irrigation jumps from drift
• Rolling 7-day precipitation = our soil water budget proxy
• Log transform + binary flag emphasise rare irrigation events

Feature Decisions

Domain-Specific Transforms

• Center each VWC depth per plot to learn deviations, not absolute bias
• Derivatives teach the LSTM to sense drying velocity
• Binary irrigation flag tells the network when actuators fired

[ _d(t) = _d(t) - _d(t-1) (t) = ((t) + 1) ]

Modeling Pipeline (Kept Simple)

• Sliding window: 168 hourly steps () 96-hour horizon
• Heavy regularised LSTM stack (512→64) to respect temporal context
• TimeSeriesSplit intended for leave-one-plot-out (bug limited generalisation – documented in findings)

Forecast Behaviour – Plot 2003

• Captures trend but under-reacts to irrigation rebounds deeper than 6"

Forecast Behaviour – Plot 2014

• Bias high at surface; deeper probes miss sharp peaks → insufficient dry-down examples

Forecast Behaviour – Plot 2015

• Tracks trend yet lags ~48 h; demonstrates training fold bug + irrigation imbalance

Injection Sensitivity Test

• Injecting +0.25 to +1.0 water after scaling failed to move predictions
• Missing feature recompute + distribution shift explain the mismatch

Honest Takeaways

• Cleaning + feature engineering are tailored to agronomic reality and reusable for next seasons
• Model failure stems from evaluation bug + post-hoc injection design, not from lack of signal
• Present the lessons learned and next steps as future work rather than unfinished business

Appendix – Swap-In Visuals

• Add actual field/sensor photography on Slide 1
• If available, include flatline sensor example or rain event on backup slide for Q&A
• Keep the PPTX (presentation/eda_story/eda_story.pptx) handy for template styling